LONG AI
Home › Blog › LongCat-2.5 Ships Today: 1.6 Trillion Pa

LongCat-2.5 Ships Today: 1.6 Trillion Parameters, Same Price

Category:资讯 · 4 min read · 2026-11-02

Today (2026-11-02) at noon, Meituan's LongCat team put LongCat-2.5 on the LongCat open platform. This isn't from a press release: the official changelog added "Version: 2026-11-02 - LongCat-2.5 live," and the pricing page now lists only 2.5 - the API is already callable.

What matters more for our readers: the plaza still sells the previous LongCat-2.0, and 2.5 isn't integrated yet.

1. What officially changed

LongCat-2.5 is positioned as a native multimodal model for long-horizon agent tasks: not text with vision bolted on, but image and text sharing the same token stream from training. The official changelog names three things - native multimodal (vision understanding fused with agent coding, enabling visual programming, creative design, 3D pages); long-horizon agent (complex codebases and full projects, multi-step reasoning, tool calling, end-to-end execution); multi-harness support (Claude Code, Codex, Hermes, with docs also covering OpenCode, Cline, Kilo Code, CodeBuddy, and WorkBuddy).

Scale per the official release: 1.6 trillion total parameters, ~48B activated per token, trained on up to 1M sequences, entirely on domestic compute; multiple outlets reported these consistently, we haven't independently re-verified. Weights aren't open yet - GitHub and HuggingFace say "coming soon" with no date.

2. Price: no hike, but marked "limited time"

The official pricing page bills per million tokens: uncached input $0.30 (¥2), cached $0.006 (¥0.04), output $1.20 (¥8). Against LongCat-2.0's official ¥2.01 / ¥8.03, 2.5 is ¥2 / ¥8 - parameters went from trillions to 1.6T plus native vision, and the unit price didn't move a cent. That's rare this year.

But the official side wrote this as a "limited-time promotional price" with no end date and no post-promo number. That's the one thing worth flagging here.

3. Which one can you actually buy here

The following is our own testing today at 13:40, no speculation. The site lists 190 models and 936 listings; the LongCat family has only 2.0:

Plaza search for longcat: longcat-2.0 card, 1.1-2.8 off, ¥0.2209/M in, ¥0.8837/M out

The math, spelled out so you can re-verify: plaza buyer price = lowest quote zhe × official CNY price. ¥2.01 × 11% ≈ ¥0.22, ¥8.03 × 11% ≈ ¥0.88, matching the page. 1.1 off is the seller-bid result and moves with supply - not a long-term promised price.

4. What it means for ordinary users

  1. On 2.0, do nothing: name, price, and availability all unchanged; the official side hasn't announced a 2.0 retirement, and 2.5 is a new model ID, not a replacement.
  2. To try 2.5 you go to longcat.chat yourself to register and top up; we can't pay on your behalf yet.
  3. No auto-upgrade: whichever model name you write in the call is the model you get.
  4. Don't treat official scores as conclusions: Terminal Bench v2.1 went 70.8→88.4 and AutomationBench 15.8→51.8, but no third-party replication and we haven't tested; in the same batch Program Bench is only 25.5, low among compared models - the official side didn't hide it.

LongCat-2.5 vs 2.0 comparison: parameters, context, official price, plaza price, availability test, and two official self-reported scores

5. Our advice

Need it now: stay on LongCat-2.0. Model name LongCat-2.0, call directly via the OpenAI-compatible API; 1.1 off is on the lower end for this tier.

Want to evaluate 2.5: run three of your own real tasks on the official side first - one code change, one image read, one long document - then compare against 2.0. Benchmarks are someone else's scenario; only you know yours.

Don't rush production traffic over: the limited-time price could end anytime; weights aren't open, so fine-tuning, quantization, and local deployment aren't ready.

Want us to list 2.5? Just ask: we'd first need to confirm upstream supply. Mention it in the group and we'll schedule by demand, then sync it to the blog and plaza once live.

Scope note

Official info from the LongCat open platform changelog (version 2026-11-02) and pricing page (longcat.chat/platform/docs/pricing/longcat-2.5), fetched 2026-11-02 13:40; parameters and benchmarks are official claims, not third-party replicated and not tested by us. Plaza price and availability are same-time API tests; prices move with seller quotes, and the final price follows the model plaza page.

相关文章

GPT-6.2 Cuts Prices 20%: What the AI Price War Means

GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer

What Is Prompt Caching? How Cache Tokens Cut API Bills

Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi

All 12 Claude Models Compared: Price and Best Use

Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload

Which AI Model Writes Code Best? Speed and Price Tested

gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi

One API Key, Every Model: How to Integrate a Multi-Model Marketplace

Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa